Investigating the applicability of a hybrid simulation methodology to motor bearing failure through incorporation of real system data

Predictive maintenance tools show promising results in the literature; however, the artificial intelligence and machine learning algorithms that underpin these approaches require large volumes of training data. This presents a challenge when sensor instrumentation is recent or limited, restricting the availability of historical data. One potential solution is the generation of simulated datasets capable of representing fault conditions. This paper investigates a hybrid simulation approach that incorporates real healthy operating data from the target system to improve simulation validity without explicitly modelling every component. The approach is evaluated using motor bearing defect experiments from Case Western Reserve University. The simulation successfully reproduces key macroscopic fault features and performs well in classification tasks with real data. However, it fails to capture the full complexity of real signatures, as confirmed by a two-group t-test comparing simulated and experimental results, highlighting limitations of simulated data substitutes.